Abstract
Modern computation based on the von Neumann architecture has become the engine that continuously fuels our modern society in many forms over the past 60 years. However, as Moore’s law is coming to an end, it is increasingly hard to continue the advancement of computing systems, calling for new technologies. Meanwhile, the rapid development of artificial intelligence (AI) and machine learning also requires powerful computing capacity as well as high power efficiency. Human brain shows extremely superior to current von-Neumann computing systems concerning low energy computation, high fault tolerance, and high efficiency in learning and decision making. Neuromorphic computing – brain-inspired computing - is becoming one of the most promising technology to advance the development of AI and computing systems after Moore’s law continuously. To mimic the way human brain carries out data-centric tasks, it adopts to implement large-scale artificial neural network (ANN) on hardware by emulation of the functions of biological neurons and synapses –two basic building blocks in the nervous system. To this goal, developing highly scalable and energy-efficient artificial neurons and synapses with bio-plausible functions is critical, but remains huge challenges because of sophisticated dynamic neural activities and the massive numbers of (1011) neurons and (1015) synapses in the human brain. Conventional complementary metal-oxide-semiconductor (CMOS) devices with binary states and complicated auxiliary circuits, cannot accommodate such requirements due to areal and energy inefficiencies. The recent advances in memristive nano-devices have opened new avenues for implementing highly dense full memristive neural networks (FMNN) comprising memristive neurons and synapses. The memristive nano-devices, which like a normal resistor regulating the flow of electricity, can additionally remember the history of charges, bringing about the ability to process and store multiple or even continuous signal states. Moreover, in combination of the crossbar structure, each memristive device can theoretically only occupy 4F2 footprint, potentially extending the artificial neural network up to human brain level. In this thesis, the novel memristive nano-devices based on new materials and device structures are explored to design the various functions of artificial neurons and synapses in an efficient way. Since memristive devices show great potential in realizing artificial synapse efficiently for neuromorphic computing, significant progress has been made on various memristive synapses to emulate rich specific synaptic functionalities. However, emulation of various subtle forms of plasticity over broad timescales in a single device remains challenges. To solve this problem, an Ag/MgO/Pt memristive device with a transient switching behavior was designed, which demonstrated various forms of synaptic plasticity from milliseconds to days in a single device. The transient behaviors are owing to the formation and spontaneous rupture of silver nano-filaments with and without electrical stimuli. By manipulating the input voltage pulse strength, such as pulse amplitude, interval, and pulse number, key features of biological synaptic plasticity including paired pulse facilitation, augmentation, post-tetanic potentiation, early-LTP and late-LTP were emulated. In addition, a reversible transition between STP and LTP was also demonstrated. The insights gained from MgO memristive devices can also aid in design of Ag/oxide/Pt memristive devices with different oxide materials to exhibit novel and diverse properties. Unlike the widely reported memristive synapses, the development of artificial neuron on memristive devices has less progress despite of its equal importance. The complicated neural dynamics is the major obstacle behind the lagging. The silver nanofilaments’ dynamics give the insight to resemble the ion movements inside the neural membrane, leading to the feasible realization of neuron’s partial essential functions: leaky integrate-and-fire which have not been reported yet. Here a rich dynamics-driven artificial neuron based on Ag/FeOx/Pt memristive device was demonstrated and it successfully emulated multiple neural features of neural processing, including leaky integration, spontaneous threshold-driven fire and self-repolarization, in a unified manner. Moreover, due to the near-zero current integration and spontaneous operations, the artificial neuron can be expected to operate at biological-level low power (10 fJ). Together with the computation role of neurons, synapses also play an essential role in neural computing by providing functions of short-term plasticity (STP) which is often underestimated. STP is the basis for synaptic computation including information processing, working memory and decision making. Although significant efforts have been made to realize synaptic devices, STP for synaptic computation by controllably modulating the synaptic strength, as well as the time constants of STP from milliseconds to minutes in a bidirectional and reversible way has remained technically challenging with conventional devices. Based on Joule heating and versatile doping induced metal-insulator transition (MIT), we firstly report synaptic computation in a single monolayer molybdenum disulfide (MoS2) devices with biologically comparable energy consumption (~tens of fJ) for per spike. A circuit with our tunable excitatory and inhibitory synaptic devices demonstrates a key function for realizing the most precise temporal computation in the human brain, sound localization: detecting an interaural time difference by suppressing sound intensity- or frequency-dependent synaptic connectivity. Although remarkable progress has been achieved in a single device, the advancements from realizing large-scale neural network comprising memristive neurons and synapses are quite slow, let alone human brain level (1011 neurons, 1015 synapses) due to fundamental challenges of integrated crossbar array structure: sneak path current and wire resistance. To overcome such issues, a self-selective memory cell based on two-dimensional (2D) hexagonal boron nitride (h-BN) and graphene in a vertical heterostructure of h-BN/graphene/h-BN was designed. By combing volatile silver filaments and non-volatile boron vacancies inside the two h-BN layers separated by impermeable graphene as a filament-blocking layer, a self-selectivity of 1010 with an on/off resistance ratio larger than 103 was demonstrated, giving the encouraging potential to realize large-scale synaptic integration. Therefore, these findings paved the way for the construction of energy-efficient large-scale FMNNs and may contribute to the development of high-density, low power, and fast neuromorphic systems.